Last [music] year, I was constrained by

token.

Now, I fixed that by joining Open AI.

Then,

I was constrained [music]

by

CPU.

And now, I feel my constraint is

actually attention.

You want to [music] optimize how often

do you need to be in the loop with your

agent.

>> [music]

>> Here we go.

>> Okay, everyone.

Don't worry, this one this one's going

to be a little bit quicker.

So, I'm Peter. Recently moved here.

Um

And you know, when you do

when you do a trending engineering,

I feel last year,

up until maybe February, I was

constrained by token.

Now, I fixed that by moving by joining

Open AI.

>> [laughter]

>> Then,

I was constrained by

CPU.

Um

I fixed that with something crap box.

I'll show you later.

And now, I feel my my constraint is

actually attention, right? So, so

all the things that you heard about like

creating loops. I mean, for me this was

like mostly a [ __ ] post and it kind of

exploded.

But what you really want to optimize

when you work

I think you Windows.

What you really want to optimize when

you work is you want to optimize how

often

do you need to be in the loop with your

agent? So, right? You need to like we

want to expand

all the things that your agent can do

from the prompt to like verification.

So, this is really

where you get more value out of it.

Also, by the way, I I forgot the ending

of the Pokémon thing. So, there's 46

endings that have exactly two Pokémon.

Uh

I don't know which one it was, but yeah,

it

It's like

well,

code fixes everything.

Um

So, I'll I'll show you like three little

things today that I think are really

cool and help our workflow.

So, number one,

you know, like I I built this thing

little thing called OpenClaw.

Uh

One of the big issues we had is

we're getting a lot of PRs, like a lot

of PRs.

And the problem is

some people literally make a PR by

typing into the coding agent

fix

and then some number

and it'll fix it and make a PR, right?

Like zero effort. And then you have a PR

that's like this long, like 4,000 lines,

and you don't really know like how much

time did that person actually Is this

even worth my time?

Uh so, one thing I ex- very method that

actually works out pretty well is

I'm going to I made a small skill.

And the skill does the following, like

after

after you create a PR, be it with Codex

or we're not discriminating here,

Claude. Um it'll ask you, "Hey, do you

want to upload the transcript? Like a a

sanitized version of the transcript."

Because

And

and let me tell you, if you do, like it

the chances that

someone looks at your PR much higher.

So, now actually we're getting a lot of

PRs. Look at this.

They like include

how people work on this.

And I find it not just me, me and also

my agent find it incredibly useful

because now I actually have a more data

points, like

did you spend 1 minute prompting this?

Did you spend 1 hour prompting this? And

like the longer your transcript is, the

higher is my confidence that you

actually care about what you fix, that

you actually understand, or at least try

to understand what you fix.

Um

and

this took me way too long because like I

was thinking too complicated. Oh, do I

need hooks? How do I build this? But

actually it's just a skill.

So I wrote this little agent transcript

skill that just basically explains

the agent how to find a transcript. I

made it so it works with every coding

agent cuz ultimately it's just JSON.

I load the JSON, it like tries its best

to sanitize it. Um I mean

ultimately trust the same agent, so it

probably won't be perfect, but you

probably don't like

don't write your

your most private secrets into the

coding session. Uh so usually it will be

fine.

And and that's it because I'm the

benevolent dictator for life, I just

made this the a default skill. So anyone

who who checks out the repository

automatically

gets the skill.

And the agent knows how to load the

skill and the and the description says,

"Load the skill when you make a PR or an

issue." Um

and bam, like our quality of PRs and

issues

for some people who do it then massively

up. Big win.

Easy to add.

Now, the second thing I want to show you

is

auto review.

Um

this is my favorite skill of all time.

Also it made everything way slower. It's

kind of like a a

continuation of what George started.

You know, whenever you you create code

it

the agent will probably do its best and

like get a version that's good.

And then you would do a PR and then like

some other agent would spin up and do a

review.

But that's slow. And that's also like

not really very practical. Um

and again it took me way too long

because I was thinking too complicated.

But really this skill again is just a

default skill and

the agent gets told load this when you

do a PR or a commit.

And then it will just invoke the CLI and

it by default it just look what's what

you have in your system and maybe maybe

codex maybe chord maybe pie maybe

anyone using open code?

But it supports open code too.

Um

and it will just invoke invoke this

skill and then

the CLI with like here review this it

will get feedback.

The beauty is

feedback goes to the coding session

where you work because that one has a

lot more context because what I

sometimes saw is

somebody would create a PR

I would load it into my codex.

Codex would do a review find three

things I fix those things but I actually

haven't fully understood what the PR is

about and all the constraints and I

actually break things

because

like from a naive standpoint yeah there

might be issues but like if you deeply

understand it maybe that's actually a

design decision right? So the beauty of

this skill is it

invokes review it gets the review back

to the original session and then the

session can decide

yeah this is right but actually no.

And the beauty of this again is if you

do that then I added instructions that

it will actually change the PR

description to include the decisions.

So then the next person who actually

looks at my PR doesn't break it again

because the naive review would find

something that you actually don't want.

Uh very easy to add again because I'm

the dictator I just made it a default.

I think I made like both anthropic and

open AI millions with just this

because it's like the massive token

burner.

Um

but you know like they have to like get

their money back at some point.

>> [laughter]

>> And uh

so and the third thing the thing that I

loved the most is

you know,

if you like

if you have these long-running sessions,

you're going to you don't want to stare

at the code. You do another session,

another session, another session. And

you have like you have like 10 coding

sessions. At some point, your it feels

like your

your your MacBook starts to become like

will take off because the CPU's are

spinning as crazy and everything become

really slow.

Um

That was seriously slowing me down. Like

at some point, I had you see like I have

multiple machines here. This is like my

computer. This is my Mac in the cloud.

This is my Mac Studio. And then this is

like I have multiple Macs where I was

like working.

But that's also kind of annoying and it

can't possibly be the perfect solution.

I felt like a crutch.

So,

cuz you keep the crustacean theme double

crap box.

Um and I actually got this idea because

I used

a blacksmith had this thing where they

created

test boxes.

But they had like daily daily outages,

so I got annoyed and I wanted to fall

back.

And then

once I had this like the affordances

kept coming.

And what it does really is it gives your

agent a little box where it can rsync

uh your changes of the repo. Like it can

actually sync it starts with the the

same state that you already have in your

GitHub Actions.

So, by default, if you have CI, which I

assume you do, it'll just work.

And then it'll rsync the changes and

then it'll run the tests in that box.

So, now no more CPU spinning because all

everything that's expensive,

it just does in that box.

And

by now it feels like it's his own

Pokémon because I support like 30

providers, um including

really weird things if you work at

companies with high security like

OpenAI.

Uh or even Parallels. So,

you will find one provider that you

like.

But then like once I had this, I was

like I was getting ideas, you know? I

was like

And initially it was just Linux. And I'm

like, ah, I have to spark, but it's like

a launch debug. It's only for Mac OS.

And this these sparks are like

especially nasty because if I test it on

my system,

my system probably is already in some

way configured and I have like

something. So, my agent will probably

break something or maybe

maybe passes because I changed something

in the past. What you really want to do

is to increase your confidence that what

you actually prompt is something that

works. You want to start with a fresh

machine.

So, I just added Windows and Linux

support.

And now

my agent has this new primitive where if

I tell him

Oops.

If I tell him, "Hey, install Open Claw

for me on

What OS is this? This is Linux.

And

Mac OS.

And Windows,

it can just do that.

And then I I gave my agent a little bit

of eyes and ears.

Uh not ears, just eyes.

Um

So, it can it can get screenshots. It

can click into it. So, I have like Chips

and Computer computer use.

So, it can actually set up your system

just as you want. And agents by now are

really good because we train them on it.

So, you can just tell him, "Hey, set up

a system where you like open

open a web browser. You preconfigure

this and this." And it'll it'll work and

like you can verify because you can see

the screens. And it might take a while,

but then at at some point you have

this thing where you can exactly on a

fresh system verify

uh your setup. And it's it's much nicer

because if you

if you would use classical computer use,

at some point you would fight with your

your cursor or like the focus would

change or like you know, it's or like

even browser users would get annoying.

And because Devon uses a fresh box, and

then you just like VNC into it,

um

you don't have any of that pain.

So, let's see if I can work this. Like

here, I have like one thing where

I was like, "This looks too boring.

Make it super gay and rainbows."

>> [laughter]

>> I should have done this before. It'll

probably take too long.

Um

or even break. Let's see. Let's see if

it manages. The The other The other

thing that that is cool on that is

um those links are shareable.

So, if I create an environment where

like, "Hey, I'm going to testing this

new feature."

I can just send this to a friend. I

mean, we we using like GitHub orgs as

the odd layer, but you can like use any

odd layer. And then they can click

around and try it, and

it's instant, right?

Um so, like I have like right now it

says, "You control."

But, I can take this link,

go on my other computer.

Oh, yeah. I also added dark and white

mode on Linux. I have no clue how it

works, but I just look for it and now

changes everything to light and dark

mode, cuz it's important when you work

all day.

Um and then like

somebody else in your org could take

control and try the new thing.

It doesn't really matter if it's slow or

fast. Like it

you'll get the job done. Like you can

imagine, hypothetically spoken,

that I work at some company where we

have a feedback channel, and like people

say,

"Wouldn't this feature be cool?" And

then like we hypothetically spoken, we

have like a

a bot in that where the bot would say,

"Yeah, let me implement this." And then

it would reply with a the

a true to a VNC session where someone

who doesn't know how to code can try out

the new feature and then realize it's a

very dumb idea.

Uh or

>> [laughter]

>> But ultimately, that's what you want,

right? You want to You want to extend

the loop. So so

any input can actually be verified. And

like And then some people realize, oh

yeah, that actually was a dumb idea. Or

some people realize, yeah, no, do it

like this. And then it was a great idea.

And then

I only need to look so like maybe a

third of the feedback things to figure

out what's actually interesting cuz it

already has been verified. Uh

and again, just by

people working in Slack using

using cloth.

Uh

so where's my rainbow?

Let's see.

Uh I guess it's still working.

Uh I should have I should have just went

with like frog green.

But yeah, those those are pretty much

it. Those are the three things that

changed my coding over the last few

months the most. We have

transcripts, which also is really great

at companies.

We have auto review.

Um

sometimes it can a little go a little

bit overboard. So you want to make sure

that

in your agents indeed, [clears throat]

you define some

some invariants. Like no, these files

will

not change after the app started. Or

like we can assume things, you know?

Like coding agents always like write for

the worst case.

And number three,

crap box.

Thanks for my TED Talk.

>> [applause]

>> Yeah.

>> The The first time we met, you told me

about your first aha moment with Open

Cloud where you sent it a voice note and

it didn't have like it found an API key

on your computer and sent you back

voices something along those lines.

>> That was scary, right?

>> And that was probably one of your like

aha moments. Have you had anything

recently that has just blown your mind

where

something really surprised you?

>> Yeah, but I can't tell you.

>> [laughter]

>> Was it stable five?

>> Uh

Fumble five? No.

>> [laughter]

>> Oh, we're getting the rainbow. Look.

Look.

Yeah, that's matching for pride.

>> So, are you using loops?

>> Um

I mean, honestly, like I think loops is

just a very very fancy work for

workflow.

So, if you create an issue now on one of

my open source repositories, an agent

will spin up

and like

verify if this is a great idea or a very

dumb idea based on the vision.md or I

just kind of define of like where I

think the project should go.

If the agent thinks it's a great idea,

it might still be a dumb idea.

But it will it will like create a PR,

another agent will spin up.

So, they're already a loop.

It's like basically it's loops and loops

because agents are already loops. But

another agent will spin up, like will

will review the PR, will like repair the

PR, make sure everything is great. So,

when I then actually have my little time

and look at the look at the open source,

I can say this, this, this and

everything's ready uh to be merged. I

don't have to

to

look 20 minutes at like a stream of

coding agents doing their thing. So,

yeah.

>> When you're designing these feedback

loops, how much uh human in the loop do

you do now versus when you first

started?

>> I mean, in the beginning I was like

kicking and screaming because the model

wouldn't do what I want when I started

like a year ago. I don't know.

The the times where I was screaming at

Claude.

Uh

And now

now it's more thinking up front.

But then it might take 1 hour, to hours

to implement and then it's basically uh

or get a look look at for example

what you can very easily implement when

you have crab box. You can create

prompts that make videos. So this is

just one feature and because videos

don't really work on GitHub we'll make a

GIF GIF I don't know. Uh and then you

basically have to be four and after

state.

And

from the prompt I have my visual

verification is like yeah, I see that

the bug here. I see that the fix here

and then press merge. So yeah, much

less.

>> At the beginning you talked about

attention being a problem. Have you

solved that?

>> I can't kill them all you know.

>> [laughter]

>> No

Well to a degree to a degree. Like I I

my agents need much less babysitting

because you just give them more tools

to do the work until at some point but

you still need

the sinking of like is this actually

something we want nobody can take away.

>> You got to be faking a transcript make

themselves sound smarter.

>> I'm honestly honestly like you know we

had this review bot

a grab tile for a while that gives the

the PR rating rating from one to five.

And we found like more than one instance

where it would give it a two and then

people just manually edited the

description to give it a five and say

this is a great PR merge. So I would not

I would not be surprised.

>> What bothers you the most about your

current workflow right now?

>> That it still requires so much sinking.

Like like we produce much more code and

we build much more complex system, which

just means that

it requires a lot more thinking. Cuz

like Asians are just not really good at

understanding how does this one thing

fit into the big picture? They focus on

the one thing. You know, they're like

this, like horses. And you're going to

push them to like look around.

But if you don't keep

the system, like how does everything fit

together? Does it actually feel right?

Yeah.

Asians can't do that yet.

All right?

>> Sweet. Thank you, Peter.

>> Thank you. [applause]